Support non-conventional optimizers

Nobody has claimed this yet.

### Bug description

I turned off the automatic optimisation, because I am using SAM optimizer (https://github.com/davda54/sam). After that, the global_step of the trainer is not updating each train step, therefore checkpointcallback are not call even though it is pass to trainer ..

used callback :

pl.callbacks.ModelCheckpoint save_weights_only=True, save_top_k=3, monitor="val_acc", mode="max", save_on_train_epoch_end=False)

### How to reproduce the bug

_No response_

### Error messages and logs

```
# Error messages and logs here please
```

### Environment

Current environment

```
#- Lightning Component (e.g. Trainer, LightningModule, LightningApp, LightningWork, LightningFlow):
- PyTorch Lightning Version 1.8.4:
- PyTorch Version 1.13:
- Python version 3.9:

```

### More info

_No response_

cc @tchaton @justusschock @awaelchli @borda @carmocca

Start with the Trainer's global_step handling when automatic optimization is disabled, then trace how ModelCheckpoint decides when to run. Reproduce the reported behavior with the SAM optimizer, Lightning 1.8.4, and the shown checkpoint settings. Done means global_step advances on each training step and the checkpoint callback is invoked as expected.

Written by the indexing model from the issue text.